Measuring health inequality among children in developing countries: does the choice of the indicator of economic status matter?

Data

The DHS program provides household-level data on health, health care utilisation and ownership of assets for about 60 low- and middle-income countries in three subsequent periods. The data are based on nationally representative surveys. In most countries, between 5,000 and 10,000 women aged 15 to 49 years were interviewed. From the Macro International web-site we obtained the most recent data (surveys held between 1991–1998) for 10 countries [17]. Bolivia, Brazil, Indonesia, Cameroon, Chad, Kenya, Malawi, Pakistan, Tanzania, and Uganda were included in the study. These countries represented a broad range of contexts in terms of region and under-5 mortality rates. They were also representative of a wider range of countries in terms of the magnitude of mortality inequalities and the pattern of under-5 mortality rates across the quintiles (with both linear and non-linear patterns included) [6]. Furthermore, we chose countries with a relatively large sample size.

Methods

When constructing an index, one of the basic decisions concerns the weights that should be given to each of the index-items. "Equal weights have the appeal of simplicity and apparent objectivity, but these qualities only mask the fact that the imposition of numeric equality is completely arbitrary" [4]. The Bank used factor scores derived through Principal Components Analysis (PCA) as the weights of the items in its asset index. Factor scores are a measure of the strength of the association of an item with the first principal component. It is assumed that this first factor represents economic status. Sample weights were not used during the PCA operation, but were used when constructing population wealth quintiles. Using this method, The Bank made for each country a separate index. In our study, we adhered to this often-used method of PCA for constructing indices.

Before we compared the World Bank (WB) index with the alternative indices, we attempted to replicate the WB index by applying PCA, using SPSS 9.0, to the same items as The Bank used. A fairly close replication of the World Bank results as reported by Gwatkin et al. [6] could be achieved. Subsequently, a series of three alternative indices was constructed by leaving out the following items: 1) all water supply and sanitation items; 2) items under 1 plus all housing items; 3) items under 1 and 2 plus electricity. In alternative Index 3 for most countries only durable consumer goods were included. The alternative indices were constructed applying the same PCA method to the shorter list of items.

The indices were compared in the following respects. First, we calculated the percentage of variance in the asset items that could be explained by the first principal component. This indicates the extent to which the variation in asset items between households can be explained by one single measure of economic status. The second feature we examined was whether it was possible to stratify the population into five, about equally large, wealth groups. Quintiles are commonly used in social epidemiology and health economics to compare health outcomes of wealth groups. It allows for a precise description of the association between wealth and health along the whole wealth gradient.

Then, we studied the changes in the categorisation of households when using the alternative indices compared to when using the WB index. Using cross tabulations we calculated the percentage of households remaining in the same quintile, the percentage that moved to the adjoining two quintiles and the percentage that moved to the furthest two quintiles.

Finally, the alternative indices were compared to the WB index in terms of their association with under-5 mortality. We calculated the under-5 mortality rates for the five wealth groups for each of the indices. Under-5 mortality was defined as the proportion of deaths under age 59 months per 1,000 live births of those born during the last 10 years preceding the survey. To have an overall measure of the magnitude of inequality across all wealth groups, we computed the Relative Index of Inequality (RII) for each of the asset indices. We compared the RII's of the alternative indices with the RII of the WB index. The RII was calculated with logistic regression using a Generalised Estimating Equations (GEE) model (using SAS 8.2) to correct for the fact that some mothers contribute more than one child to the set of observations, and that the children born to one mother are not statistically independent. The RII is a summary measure of the differences in mortality experience between all five wealth groups, and can be interpreted as the (estimated) ratio in odds between the poorest and the richest household. This estimate captures only the linear relationship between wealth and health [18].

We performed analogous analyses for coverage rates in measles immunisation. For pragmatic reasons we used the rate of children not immunised against measles as outcome measure. The use of this negatively stated measure facilitates comparison with under-5 mortality, also a negatively stated indicator. The rate of children not immunised against measles, here called "immunisation coverage" in short, is defined as the rate of children aged 12–23 months during the survey who were not immunised against measles. We calculated and compared this rate for all five wealth groups using each of the asset indices. We also calculated and compared the RII.

When examining sensitivity, a problem arises if the sizes of the wealth groups differ between the indices. The comparison between indices of group-specific rates is only possible when the groups that are compared are of the same size. If the group size is different, the change in rate could be an artefact, and conclusions can be stated with less certainty. When comparing the RII's between the indices, different group sizes are generally not a problem as this measure takes into account the size and relative position of the wealth groups.

Explaining the results

Sensitivity of inequality to alternative constructions of the WB index is likely to be related to the low common variance of the items in the WB index. This low common variance can be explained by the fact that the WB includes a broad range of different items, each of which has its own determinants besides economic status. Upon exclusion of items from this index, the common variance increased. The reason is that the new lists were shorter and consisted of more homogeneous sets of items. As a consequence, the categorisation of households into wealth groups changed, leading in its turn to different mortality rates per wealth group.

Even though the observed sensitivity is understandable, it is more puzzling why the use of alternative indices had different effects for different countries as well as for different outcome indicators. Below, we will forward some explanations.

We hypothesised that inequality in under-5 mortality would decrease upon exclusion of water and sanitation items from the WB index and would further decrease upon exclusion of housing items. This is because we expected that part of the relationship between wealth as measured by the WB index and under-5 mortality would be explained by variables that have a direct effect on child mortality, apart from their indirect effect as indicators of economic status [7]. For a number of countries we saw this expected decrease. This supports the hypothesis that for some countries part of inequality in mortality measured using the WB index can be attributed to direct determinants of health rather then to economic status alone.

This hypothesis, however, cannot explain the decrease in inequality in immunisation coverage observed for some countries upon exclusion of direct determinants of health from the index. The reason is that housing characteristics, and water and sanitation facilities only influence immunisation coverage as indicators of economic status, and don't have a 'direct' impact that is comparable to their effect on mortality. An alternative hypothesis would be that water and sanitation facilities and housing characteristics are also indicators for regional development or rural/urban residence. For instance, using the bush as latrine probably indicates rural residence, whereas using a private toilet is probably more related to urban residence. Therefore, the decrease in observed inequality in both mortality and immunisation may also in part be explained by the fact that the WB index captures rural-urban differences in both wealth and the health indicators.

We expected a further decrease in inequality upon subsequent exclusion of electricity from the index. Electricity can be an indicator of community wealth. Regional disparities in the availability of electricity probably run parallel to disparities in access to and quality of health care services and disparities in mortality. When excluding electricity from the asset index, these regional disparities in wealth and mortality as measured through electricity, are given less weight. Doing so, one can expect a decrease in inequality in mortality. We saw that indeed for a number of the countries, inequality in under-5 mortality decreased upon exclusion of electricity from the wealth index. This may indicate that health inequality as measured by the WB index, through electricity, also captures for some countries some of the regional disparities in wealth and health.

The hypotheses above, however, cannot explain why in some cases inequality was not sensitive, and why in one case inequality in under-5 mortality increased, instead of decreased. Additional explanations therefore need to be sought.

Inequality in under-5 mortality was robust to changes in the measure of economic status used for Uganda. This is related to the fact that Uganda was the only country for which items on food sufficiency were included in the WB index. When doing an additional analysis, excluding food sufficiency from all four indices, the RII became slightly sensitive also for this country (the largest change in RII being 14%, from 1.68 when using the WB index to 1.77 when using Index 1). Sensitivity thus may depend on the specific items included in the asset index.

The slight increase in inequality in under-5 mortality in Cameroon upon exclusion of water and sanitation items could not be attributed to the above factors. As already mentioned, household ownership of assets is also determined by other factors besides economic status, such as local availability and preferences. These factors can act as confounders in the relationship between household wealth and child mortality. Apparently, these confounders are in some cases more difficult to disentangle than in others. Multivariate analysis would be needed to gain more insight into these relationships, and thus into the underlying mechanisms linking wealth and health.

Implications

Our study shows that researchers and policy makers should be aware that the choice between alternative indicators of economic status often does affect, and in some cases to an important extent, the observed magnitude of poor-rich differences in health and health-related outcome measures. It also shows that it is difficult to predict the size and direction of sensitivity.

This is important in the present context in which monitoring and tackling poor-rich inequalities in developing countries have become increasingly important policy objectives, and in which many studies are being published on this issue. One of the major difficulties this new field is facing, is determining who is rich and who is poor. An index based on household ownership of assets is an often-used way to do so. Different researchers, however, use different sets of asset items. Our study shows that who is defined as poor and who as rich, varies with the asset items included in the index.

Our study implies that we should be extremely careful comparing results of studies using different indicators of economic status, as differences between countries and trends over time may in part be an artefact of the different indicators used. This is important both for monitoring health inequalities, evaluating the effects of policy interventions on these inequalities, and for targeting the poor in health policies. The choice of the measure of economic status should therefore be carefully made.

For descriptive and monitoring purposes we advise to use a comprehensive list of asset items such as used by the World Bank. A good alternative would be a much more extended list of consumer goods. In countries or regions where durable consumer goods are hardly accessible to anyone, or where investments in housing and amenities are given priority, the latter can be important indicators of economic status or wealth. Moreover, surveys such as the DHS only include a limited number of durable consumer goods, whereas items that the poor and inhabitants of rural areas are likely to own (e.g. chair, plastic recipients, animals, farming tools) are not included. In such cases, the inclusion of water, sanitation and housing items facilitates stratification of households at the lower end of the wealth ladder. For these reasons, it would generally be advisable to use a comprehensive list of asset items for descriptive and monitoring studies.

For explanatory studies, though, it can be important to analyse the different sets of asset items separately, and not to combine them into one index. It enables the assessment of the relative importance of different components of material wealth, especially water and sanitation versus housing versus consumer items versus indicators of community wealth. Estimates of the relative importance of these components can contribute to the detection of causal mechanims that are most responsible for high child mortality among poor families. This information is important for intervention purposes, since it addresses questions such as: would it be more effective to invest in income generating projects or in housing, water and sanitation programs; and should development efforts be focussed on the household level or the community level? For such explanatory studies it would be advisable to use multiple regression, path analysis or similar multivariate techniques.

For those designing new surveys intending to measure economic status in developing countries, we advice to also include items that poor households are likely to own and indicators of economic status in rural areas, such as the ownership of land, animals, and farming tools. Also the inclusion of context-specific indicators of economic status, as shown by the example of 'food sufficiency' in Uganda, would be useful when aiming to make a refined stratification along the lines of economic status. The inclusion of 'rural' and context specific items can also be important for making a proper identification of target groups for health policies.